PROJECT SUMMARY/ABSTRACT Candidate: Atalie Carina Thompson, MD, MPH is a current glaucoma fellow and Heed fellow with a long-term career goal of becoming an independent clinician-scientist and leader in the field of glaucoma and public health. She has a long-standing interest in addressing healthcare disparities in medicine, and in improving the diagnosis of glaucoma and other ophthalmic diseases through imaging technology. While obtaining a medical degree at Stanford, she received a fellowship to complete a master?s degree in public health with additional higher-level coursework in biostatistics and epidemiology. Her immediate goal in this proposal is to refine and validate a deep learning (DL) algorithm capable of quantifying neuroretinal damage on optic disc photographs and then to apply it in a pilot teleophthalmology program. With a K23 Mentored Patient-Oriented Research Career Development Award, she will acquire additional didactic training and mentored research experience in glaucoma imaging, machine learning, biostatistics, clinical research, and the responsible conduct of research. Environment: The mentorship and expertise of the advisory committee, the extensive resources at the Duke Eye Center and Departments of Biostatistics and Biomedical Engineering, and the significant institutional commitment will provide her with the support needed to transition successfully into an independent clinician-scientist. Research: This proposal will test the hypothesis that a DL algorithm trained with SDOCT detects glaucoma on optic disc photographs with greater accuracy than human graders. In Specific Aim 1, a DL algorithm that quantifies neuroretinal damage on optic disc photographs will be refined. The main hypothesis is that the quantitative output provided by the DL algorithm will allow accurate discrimination of eyes at different stages of the disease according to standard automated perimetry, and will generate cut-offs suitable for use in a screening setting. In Specific Aim 2, the short-term repeatability and reproducibility of the DL algorithm in optic disc photographs acquired over a time period of several weeks will be determined. The hypothesis is that the test-retest variability of the predictions from the DL algorithm will be similar to the original measurements acquired by SDOCT. In Specific Aim 3, the DL algorithm will be applied to optic disc photographs obtained during a pilot screening teleophthalmology program in primary care clinics and assisted living facilities. The hypothesis is that the DL algorithm will be more accurate than human graders when a full ophthalmic examination is used as the gold standard. This work will constitute the basis of an R01 grant and will advance our understanding of the application of deep learning algorithms in glaucoma and teleophthalmology.